Hi all, I tried to summarize below all approaches we were discussing.
In case there are some difficulties in email formatting, see GoogleDocs: https://docs.google.com/document/d/1RjlcebpigOj9DyLCedyRxki9nZcFdkBchy-k7BiThMc/edit?usp=sharing <https://docs.google.com/document/d/1RjlcebpigOj9DyLCedyRxki9nZcFdkBchy-k7BiThMc/edit?usp=sharing> Personally, I find options A01 and A02/B03 very convenient. Please, let me know what you think and feel free to add more options. Thanks, Konstantin ––––––––––––––––––––––––––––––––––––– Preconditions: we would need to remote negative timestamp check from producer anyway no existing data should have -1 timestamp unless it is NO_TIMESTAMP older client are not embedding a timestamp: default to -1 on the broker side... APPEND_TIME would still work for older producers Can we still use −1 as NO_TIMESTAMP value? Yes: A01 Keep using −1ms as a NO_TIMESTAMP value No, A02 We want to change the semantics of −1ms as an actual timestamp. For example, use Long.MIN_VALUE as "NO_TIMESTAMP". How do we indicate new timestamp format? A02 In the message itself. That way one topic may contain messages of different types. How? B01 Create new timestamp type and use one of the reserved bits B02 Use one reserved bit in the message as a flag B03 Introduce new message format version By the topic property: `allow.negative.timestamp`. That way messages with CreateTime could have negative value. "false" by default for existing topics. Should `allow.negative.timestamp` be enabled by default for new topics? Yes. C01: We want it to be future standard No. C02: We want it to be client specific case by case decision. Do we allow changing that property for existing topics? Yes Yes, but only if existing topic have no records with negative timestamps already No: old behavior and new behavior are not mixed Can old producers write to topics with that flag enabled? Yes, as long as we can differentiate old and new messages: B01-03 No, broker will write old "NO_TIMESTAMP" as a −1ms No, broker will throw an error. By the broker config: `allow.negative.timestamp`. What to do with old topics? D01: Convert all messages in all topics to the new format (B01-03) during the update phase. D02: Keep old messages, but write new messages in the new format. D03: Keep old messages, but new topics will have messages with new format. A01 Keep using −1ms as a NO_TIMESTAMP value Cost of change: minimal Pros Cons Realistically, the scope of impact for that -1 is pretty tiny, as has been pointed out. A single millisecond gap in available timestamps in 1969. Keeping -1 as "unknown" leaving a time gap gives a lot of headaches Streams drops records with timestamp -1 but processes other records with negative timestamps which may cause a number of question on the mailing list and will confuse users that don't pay attention and "loose" data. A02 Indicate that message may have negative timestamp in the message itself Cost of change: moderate A03 Add topic config that allows messages with negative timestamp Cost of change: high Kafka Streams could check the topic config and only drop negative timestamps is they are not enabled Pros Cons Relaxes compatibility concerns We already have way too many configs Introducing the new config seems like it has significant compatibility concerns that need to be sorted out. A04 Add broker config that allows messages with negative timestamp Cost of change: high – B01 Add new timestamp type that can be negative Cost of change: moderate Pros Cons Changing message format might not be worth it B02 Use one of reserved bit to indicate negative timestamp (aka flag) Cost of change: moderate Pros Cons Using a flag bit might not be worth it B03 Introduce new message format version Cost of change: low Old producers will still be able to write messages in old version Pros Cons Clear semantics: in new version CreateTime can be negative. Using a flag bit might not be worth it > On Jan 2, 2018, at 11:26 PM, Ewen Cheslack-Postava <e...@confluent.io> wrote: > > On Tue, Jan 2, 2018 at 8:04 PM, Matthias J. Sax <matth...@confluent.io> > wrote: > >> I was thinking about a broker/topic config. >> >> However, I am not sure if we only need to worry about data written in >> the future (this would only be true, if there would be no records with >> -1 timestamp already). Assume that we you an existing topic that >> contains data with -1 = UNKNOWN records -- for this case, we would give >> those timestamps a new semantics if we suddenly allow negative >> timestamps. (Assuming that we don't allow -1 as a gap in the timeline >> what I would rather not do.) >> > > Using the Java producer you cannot have a negative timestamp today. So > (modulo comment about being dependent on the client implementation), no > existing data should have -1 timestamp unless it is NO_TIMESTAMP. > > When you say you'd rather not like to have -1 as a gap in the timeline, can > you explain the potential scale of impact? I view it as a relatively > unlikely value and something that people who are really concerned with > negative timestamps can easily work around. Probably many users won't care > as they will not be using pre-1970s data where they actually set the Kafka > timestamp (rather than having timestamps embedded in the data) anyway. I > agree it isn't ideal, but to me it looks like a reasonable tradeoff. What > are the effects/use cases that make you concerned that we'd see significant > user pain as a result? > > >> >> Also note, that it's not really client specific IMHO, as one could >> implement their own clients. There are many third party clients and we >> don't know if they check for negative timestamps (applications could >> even assign their own special meaning to negative timestamps as those >> are unused atm) -- furthermore, all older client not embedding a >> timestamp default to -1 on the broker side... >> > > I said "client-specific" because some of the checks are done on the > client-side, which means they are dependent on the specific client > implementation being used. Based on the rest of your comment, I think we're > in agreement except for how we are naming things :) I'd have to double > check if the same level of enforcement is done broker-side. I only mention > that because we tend to discuss these proposals in the context of only the > Java clients, but it is worth thinking through the impact to other clients > as well. > > >> >>> The implementation could easily be made to map >>> those values into a range that is less likely to be utilized (e.g. use >> the >>> values near Long.MIN_VALUE and have the consumer convert back as needed). >>> The sentinel for NO_TIMESTAMP could be changed between versions as long >> as >>> it is handled consistently between client versions. >> >> This opens Pandora's box IMHO. >> > > Why? There should be a small number of values that need to be mapped and > someone could think through the different compatibility issues that are > possible to determine if there are any significant issues/drawbacks. > > >> >>> Introducing the new >>> config seems like it has significant compatibility concerns that need to >> be >>> sorted out. >> >> I cannot follow here -- from my point of view, it relaxes compatibility >> concerns. If we only allow new topic to enable negative timestamps, old >> behavior and new behavior are not mixed. IMHO, mixing both would be a >> real issue. Thus, for new topics we can change "unknown" from -1 to >> Long.MIN_VALUE and don't mix two different approaches within a single >> topic. >> > > What's the mechanism for this? Is the new config only allowed in > CreateTopics requests? If you use existing tooling to set topic configs, > you would just be able to set any valid config. Are the semantics just > undefined if you do? Unless it is impossible to do certain things, we have > to deal with the compatibility concerns regardless of intended use. Might > be fine to just say the behavior is undefined, but there's still work to be > done there. Regardless, I didn't (and probably still don't) have a concrete > understanding of the proposed setting, so hard for me to reason about it. > > -Ewen > > >> >> I see your point that we do have too many configs -- we could also make >> it a new value for existing `message.timestamp.type`. >> >> >> -Matthias >> >> >> On 1/2/18 7:48 PM, Ewen Cheslack-Postava wrote: >>> For `allow.negative.timestamps`, do you mean this as a broker config? I'm >>> not entirely clear on what the proposal would entail. >>> >>> I think taking into account whether we're talking about compatibility >> with >>> existing data in Kafka vs enabling use of negative timestamps is >> important >>> here. If they're effectively not supported today (though admittedly this >> is >>> really client-specific), then we need only concern ourselves with data >> that >>> hasn't been produced into Kafka yet. In that case, we can always handle >>> sentinel values in special ways if we really want to. For example, the >> Java >>> producer does not accept any values < 0 and the API supports passing null >>> rather than the sentinels. The implementation could easily be made to map >>> those values into a range that is less likely to be utilized (e.g. use >> the >>> values near Long.MIN_VALUE and have the consumer convert back as needed). >>> The sentinel for NO_TIMESTAMP could be changed between versions as long >> as >>> it is handled consistently between client versions. >>> >>> IMO we already have way too many configs, so we should think about where >>> the impact is and if a not ideal, but also not significant compromise can >>> be made and avoid most of the additional complexity. Introducing the new >>> config seems like it has significant compatibility concerns that need to >> be >>> sorted out. In contrast, I suspect the use cases we need to support that >>> have come up so far can handle 1 or 2 special cases and the necessary >>> munging could be handled safely by interceptors such that it is trivial >> to >>> make sure all your apps do the right thing. I appreciate the pain of a >> ton >>> of mailing list questions about an issue like this, but given the >>> likelihood of encountering that particular value, I just find it unlikely >>> it would be that common and I think it's a reasonable tradeoff to tell a >>> user they might need to handle that one special case. >>> >>> -Ewen >>> >>> On Thu, Dec 28, 2017 at 12:58 PM, Matthias J. Sax <matth...@confluent.io >>> >>> wrote: >>> >>>> I agree that changing message format or using a flag bit might not be >>>> worth it. >>>> >>>> However, just keeping -1 as "unknown" leaving a time gap give me a lot >>>> of headache, too. Your arguments about "not an issue in practice" kinda >>>> make sense to me, but I see the number of question on the mailing list >>>> already if we really follow this path... It will confuse users that >>>> don't pay attention and "loose" data if Kafka Streams drops records with >>>> timestamp -1 but processes other records with negative timestamps. >>>> >>>> Thus, I was wondering if a new topic config (maybe >>>> `allow.negative.timestamps` with default `false`) that allows for enable >>>> negative timestamps would be the better solution? With this new config, >>>> we would not have any sentinel value for "unknown" and all timestamps >>>> would be valid. Old producers, can't write to those topics if they are >>>> configured with CREATE_TIME though; APPEND_TIME would still work for >>>> older producers but with APPEND_TIME no negative timestamps are possible >>>> in the first place, so this config would not have any impact anyway. >>>> >>>> Kafka Streams could check the topic config and only drop negative >>>> timestamps is they are not enabled. Or course, existing topic should not >>>> enable negative timestamps if there are records with -1 in them already >>>> -- otherwise, semantics break down -- but this would be a config error >>>> we cannot prevent. However, I would expect that mostly newly created >>>> topics would enable this config anyway. >>>> >>>> >>>> -Matthias >>>> >>>> On 12/18/17 10:47 PM, Ewen Cheslack-Postava wrote: >>>>> I think the trivial change of just recognizing using -1 was a mistake >>>> for a >>>>> sentinel value and special casing it while allowing other negative >> values >>>>> through is the most practical, reasonable change. >>>>> >>>>> Realistically, the scope of impact for that -1 is pretty tiny, as has >>>> been >>>>> pointed out. A single millisecond gap in available timestamps in 1969. >>>> For >>>>> producers that really want to be careful (as the NYT data might want to >>>>> be), having the producer layer adjust accordingly is unlikely to be an >>>>> issue (you can't assume these timestamps are unique anyway, so they >>>> cannot >>>>> reasonably used for ordering; adjusting by 1ms is a practical >> tradeoff). >>>>> >>>>> Other approaches where we modify the semantics of the timestamp from >> the >>>>> two existing modes require eating up valuable flags in the message >>>> format, >>>>> or ramping the message format version, all of which make things >>>>> significantly messier. Hell, timezones, leap seconds, and ms >> granularity >>>>> probably make that 1ms window pretty much moot for any practical >>>>> applications, and for the extremely rare case that an application might >>>>> care, they are probably willing to pay the cost of a secondary index if >>>>> they needed to store timestamp values in the payload rather than in the >>>>> metadata. >>>>> >>>>> Given that we have the current system in place, I suspect that any >>>>> translation to using Long.MIN_VALUE as the sentinel is probably just >> more >>>>> confusing to users, adds more implementation overhead to client >>>> libraries, >>>>> and is more likely to introduce bugs. >>>>> >>>>> Warts like these always feel wrong when approached from pure design >>>>> principles, but the fact is that the constraints are already there. To >>>> me, >>>>> none of the proposals to move to an encoding we'd prefer seem to add >>>> enough >>>>> value to outweigh the migration, compatibility, and implementation >> costs. >>>>> >>>>> @Dong -- your point about special timestamp values is a very good one. >>>> The >>>>> issue may extend to other cases in the protocol where we use >> timestamps. >>>> Is >>>>> this the scope we need to worry about (2 values instead of just 1) or >> are >>>>> there others? This also might be something we want to look out for in >> the >>>>> future -- using special values relative to <SignedIntType>.MIN_VALUE >>>>> instead of relative to 0. >>>>> >>>>> -Ewen >>>>> >>>>> On Tue, Dec 12, 2017 at 11:12 AM, Dong Lin <lindon...@gmail.com> >> wrote: >>>>> >>>>>> Hey Konstantin, >>>>>> >>>>>> Thanks for updating the KIP. >>>>>> >>>>>> If we were to support negative timestamp in the message, we probably >>>> also >>>>>> want to support negative timestamp in ListOffsetRequest. Currently in >>>>>> ListOffsetRequest, timestamp value -2 is used to indicate earliest >>>>>> timestamp and timestamp value -1 is used to indicate latest timestamp. >>>> It >>>>>> seems that we should make changes accordingly so that -1 and -2 can be >>>>>> supported as valid timestamp in ListOffsetRequest. What do you think? >>>>>> >>>>>> Thanks, >>>>>> Dong >>>>>> >>>>>> >>>>>> >>>>>> On Mon, Dec 11, 2017 at 12:55 PM, Konstantin Chukhlomin < >>>>>> chuhlo...@gmail.com >>>>>>> wrote: >>>>>> >>>>>>> Hi all, >>>>>>> >>>>>>> I've updated KIP with few more details: >>>>>>> Added (proposed) Changes in binary message format < >>>>>>> https://cwiki.apache.org/confluence/display/KAFKA/KIP-228+ >>>>>>> Negative+record+timestamp+support#KIP-228Negativerecordtimes >>>>>>> tampsupport-Changesinbinarymessageformat> >>>>>>> Added Changes from producer perspective < >> https://cwiki.apache.org/conf >>>>>>> luence/display/KAFKA/KIP-228+Negative+record+timestamp+supp >>>>>>> ort#KIP-228Negativerecordtimestampsupport- >>>> Changesfromproducerperspective >>>>>>> >>>>>>> Added Changes from consumer perspective < >> https://cwiki.apache.org/conf >>>>>>> luence/display/KAFKA/KIP-228+Negative+record+timestamp+supp >>>>>>> ort#KIP-228Negativerecordtimestampsupport- >>>> Changesfromconsumerperspective >>>>>>> >>>>>>> >>>>>>> Let me know if it makes sense to you. >>>>>>> >>>>>>> -Konstantin >>>>>>> >>>>>>>> On Dec 7, 2017, at 2:46 PM, Konstantin Chukhlomin < >>>> chuhlo...@gmail.com >>>>>>> >>>>>>> wrote: >>>>>>>> >>>>>>>> Hi Matthias, >>>>>>>> >>>>>>>> Indeed for consumers it will be not obvious what −1 means: actual >>>>>>> timestamp >>>>>>>> or no timestamp. Nevertheless, it's just −1 millisecond, so I >> thought >>>>>> it >>>>>>> will be >>>>>>>> not a big deal to leave it (not clean, but acceptable). >>>>>>>> >>>>>>>> I agree that it will much cleaner to have a different type of topics >>>>>>> that support >>>>>>>> negative timestamp and/or threat Long.MIN_VALUE as a no-timestamp. >>>>>>>> I'll update KIP to make it a proposed solution. >>>>>>>> >>>>>>>> Thanks, >>>>>>>> Konstantin >>>>>>>> >>>>>>>>> On Dec 5, 2017, at 7:06 PM, Matthias J. Sax <matth...@confluent.io >>> >>>>>>> wrote: >>>>>>>>> >>>>>>>>> Thanks for the KIP Konstantin. >>>>>>>>> >>>>>>>>> From my understanding, you propose to just remove the negative >>>>>> timestamp >>>>>>>>> check in KafkaProducer and KafkaStreams. If topics are configured >>>> with >>>>>>>>> `CreateTime` brokers also write negative timestamps if they are >>>>>> embedded >>>>>>>>> in the message. >>>>>>>>> >>>>>>>>> However, I am not sure about the overlapping semantics for -1 >>>>>> timestamp. >>>>>>>>> My concerns is, that this ambiguity might result in issues. Assume >>>>>> that >>>>>>>>> there is a topic (configured with `CreateTime`) for which an old >> and >>>> a >>>>>>>>> new producer are writing. The old producer uses old message format >>>> and >>>>>>>>> does not include any timestamp in the message. The broker will >>>>>> "upgrade" >>>>>>>>> this message to the new format and set -1. At the same time, the >> new >>>>>>>>> producer could write a message with valid timestamp -1. A consumer >>>>>> could >>>>>>>>> not distinguish between both cases... >>>>>>>>> >>>>>>>>> Also, there might be other Producer implementations that write >>>>>> negative >>>>>>>>> timestamps. Thus, those might already exist. For Streams, we don't >>>>>>>>> process those and we should make sure to keep it this way (to avoid >>>>>>>>> ambiguity). >>>>>>>>> >>>>>>>>> Thus, it might actually make sense to introduce a new timestamp >> type >>>>>> to >>>>>>>>> express those new semantics. The question is still, how to deal >> with >>>>>>>>> older producer clients that want to write to those topics. >>>>>>>>> >>>>>>>>> - We could either use `Long.MIN_VALUE` as "unknown" (this would be >>>> way >>>>>>>>> better than -1 as it's not in the middle of the range but at the >> very >>>>>>>>> end and it will also have well-defined semantics). >>>>>>>>> - Or we use a "mixed-mode" where we use broker wall-clock time for >>>>>>>>> older message formats (ie, append time semantics for older >> producers) >>>>>>>>> - Third, we would even give an error message back to older >> producers; >>>>>>>>> this might change the backward compatibility guarantees Kafka >>>> provides >>>>>>>>> so far when upgrading brokers. However, this would not affect >> exiting >>>>>>>>> topics, but only newly created ones (and we could disallow changing >>>>>> the >>>>>>>>> semantics to the new timestamp type to guard against miss >>>>>>>>> configuration). Thus, it might be ok. >>>>>>>>> >>>>>>>>> For Streams, we could check the topic config and process negative >>>>>>>>> timestamps only if the topic is configures with the new timestamp >>>>>> type. >>>>>>>>> >>>>>>>>> >>>>>>>>> Maybe I am a little bit to paranoid about overloading -1 semantics. >>>>>>>>> Curious to get feedback from others. >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> -Matthias >>>>>>>>> >>>>>>>>> >>>>>>>>> On 12/5/17 1:24 PM, Konstantin Chukhlomin wrote: >>>>>>>>>> Hi Dong, >>>>>>>>>> >>>>>>>>>> Currently we are storing historical timestamp in the message. >>>>>>>>>> >>>>>>>>>> What we are trying to achieve is to make it possible to do Kafka >>>>>> lookup >>>>>>>>>> by timestamp. Ideally I would do `offsetsForTimes` to find >> articles >>>>>>> published >>>>>>>>>> in 1910s (if we are storing articles on the log). >>>>>>>>>> >>>>>>>>>> So first two suggestions aren't really covering our use-case. >>>>>>>>>> >>>>>>>>>> We could create a new timestamp type like "HistoricalTimestamp" or >>>>>>> "MaybeNegativeTimestamp". >>>>>>>>>> And the only difference between this one and CreateTime is that it >>>>>>> could be negative. >>>>>>>>>> I tend to use CreateTime for this purpose because it's easier to >>>>>>> understand from >>>>>>>>>> user perspective as a timestamp which publisher can set. >>>>>>>>>> >>>>>>>>>> Thanks, >>>>>>>>>> Konstantin >>>>>>>>>> >>>>>>>>>>> On Dec 5, 2017, at 3:47 PM, Dong Lin <lindon...@gmail.com> >> wrote: >>>>>>>>>>> >>>>>>>>>>> Hey Konstantin, >>>>>>>>>>> >>>>>>>>>>> Thanks for the KIP. I have a few questions below. >>>>>>>>>>> >>>>>>>>>>> Strictly speaking Kafka actually allows you to store historical >>>>>> data. >>>>>>> And >>>>>>>>>>> user are free to encode arbitrary timestamp field in their Kafka >>>>>>> message. >>>>>>>>>>> For example, your Kafka message can currently have Json or Avro >>>>>>> format and >>>>>>>>>>> you can put a timestamp field there. Do you think that could >>>> address >>>>>>> your >>>>>>>>>>> use-case? >>>>>>>>>>> >>>>>>>>>>> Alternatively, KIP-82 introduced Record Header in Kafka and you >> can >>>>>>> also >>>>>>>>>>> define your customized key/value pair in the header. Do you think >>>>>>> this can >>>>>>>>>>> address your use-case? >>>>>>>>>>> >>>>>>>>>>> Also, currently there are two types of timestamp according to >>>>>> KIP-32. >>>>>>> If >>>>>>>>>>> the type is LogAppendTime then the timestamp value is the time >> when >>>>>>> broker >>>>>>>>>>> receives the message. If the type is CreateTime then the >> timestamp >>>>>>> value is >>>>>>>>>>> determined when producer produces message. With these two >>>>>>> definitions, the >>>>>>>>>>> timestamp should always be positive. We probably need a new type >>>>>> here >>>>>>> if we >>>>>>>>>>> can not put timestamp in the Record Header or the message >> payload. >>>>>>> Does >>>>>>>>>>> this sound reasonable? >>>>>>>>>>> >>>>>>>>>>> Thanks, >>>>>>>>>>> Dong >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> On Tue, Dec 5, 2017 at 8:40 AM, Konstantin Chukhlomin < >>>>>>> chuhlo...@gmail.com> >>>>>>>>>>> wrote: >>>>>>>>>>> >>>>>>>>>>>> Hi all, >>>>>>>>>>>> >>>>>>>>>>>> I have created a KIP to support negative timestamp: >>>>>>>>>>>> https://cwiki.apache.org/confluence/display/KAFKA/KIP- >>>>>>>>>>>> 228+Negative+record+timestamp+support < >> https://cwiki.apache.org/ >>>>>>>>>>>> confluence/display/KAFKA/KIP-228+Negative+record+timestamp+ >>>>>> support> >>>>>>>>>>>> >>>>>>>>>>>> Here are proposed changes: https://github.com/apache/ >>>>>>>>>>>> kafka/compare/trunk...chuhlomin:trunk < >> https://github.com/apache/ >>>>>>>>>>>> kafka/compare/trunk...chuhlomin:trunk> >>>>>>>>>>>> >>>>>>>>>>>> I'm pretty sure that not cases are covered, so comments and >>>>>>> suggestions >>>>>>>>>>>> are welcome. >>>>>>>>>>>> >>>>>>>>>>>> Thank you, >>>>>>>>>>>> Konstantin >>>>>>>>>> >>>>>>>>> >>>>>>>> >>>>>>> >>>>>>> >>>>>> >>>>> >>>> >>>> >>> >> >>
